摘要
最小二乘逆时偏移(LSRTM)通常基于梯度类算法,经过几十次甚至上百次的迭代得到最终的成像剖面,然而常规最小二乘逆时偏移其在迭代过程中,所求梯度通常不做优化处理,导致最小二乘逆时偏移的收敛效率和成像精度不高,并且每次迭代的模型更新处理还需付出1~2次的波场延拓计算代价来获取迭代步长.本文将深度学习中的优化算法QHAdam引入到传统时间域最小二乘逆时偏移计算中,可在付出极小计算代价的前提下,直接获得优化的模型更新量,同时避免了迭代步长的求取.Marmousi模型实验结果显示,相比于常规最小二乘逆时偏移算法,基于QHAdam梯度优化算法的最小二乘逆时偏移其收敛效率和成像精度更高,且由于减少了迭代步长的求取步骤,其也具有更高的计算效率;相对于基于Adam算法最小二乘逆时偏移,本文方法也具有更高的收敛效率和收敛精度.
High-resolution imaging result derived by Least-Squares Reverse Time Migration(LSRTM)is achieved through iterative gradient calculations and model updates usually.But in conventional LSRTM,the gradient without optimization will lead to low convergence efficiency and low imaging accuracy.Besides,the calculation of step-length in iteration introduces extra wave-field extrapolation.This paper introduces a deep learning optimization algorithm,called QHAdam,into conventional time-domain LSRTM,optimizing the gradients and calculating variations of models directly with less computational cost.Experiments based on Marmousi model show that the LSRTM based on QHAdam method has higher computational efficiency compared with conventional algorithm since avoiding the calculation of step-lengths for every iteration except for the first one,and it also has higher convergence efficiency and imaging accuracy than that based on conventional and Adam algorithms.
作者
王绍文
宋鹏
谭军
解闯
赵波
毛士博
WANG ShaoWen;SONG Peng;TAN Jun;XIE Chuang;ZHAO Bo;MAO ShiBo(College of Marine Geo-sciences,Ocean University of China,Qingdao 266100,China;Laboratory for Marine Mineral Resource,Pilot National Laboratory for Marine Science and Technology,Qingdao 266100,China;Key Laboratory of Submarine Geosciences and Prospecting Techniques Ministry of Education,Qingdao 266100,China)
出处
《地球物理学报》
SCIE
EI
CAS
CSCD
北大核心
2022年第7期2673-2680,共8页
Chinese Journal of Geophysics
基金
青岛海洋科学与技术试点国家实验室山东省专项经费“问海计划”项目(2021WHZZB0700)
国家自然科学基金项目(42074138)
中央高校基本科研业务费专项(201964016)
山东省重大科技创新工程项目(2019JZZY010803)联合资助。
关键词
最小二乘逆时偏移
梯度优化
QHAdam
深度学习
Least-Squares Reverse Time Migration(LSRTM)
Gradient optimization
QHAdam
Deep learning
